LATCHBIO

OLYMPUS opened an institutional file on LatchBio because its name is in two labs’ release documents for opposite reasons. xAI’s Grok 4.7 post says its model tops “LatchBio’s biosafety benchmark at 62.4%” (xAI, Introducing Grok 4.7, 21 Sep 2026), and six Anthropic system cards use LatchBio’s bioinformatics evaluations as capability evidence, while LatchBio’s own blog reports that Anthropic’s Fable 5.1 refused every long-horizon task it was given (LatchBio blog, 1 Sep 2026). This is not a psychometric profile, since an institution has no Dark Triad. It is a mandate, an investor list, a client claim, a set of lab documents and a voice, and the finding is the shape of the arrangement: a venture-backed biology-software company has become the benchmark labs cite in their own safety documents, and it also sells the audits.

Evaluator record: the engagements, the terms each document states, the findings quoted and what followed.

Institutional Archetype

THE BENCHMARK THAT GRADES THE LABS AND SELLS THEM THE AUDIT — LatchBio’s biosecurity page: “Helping frontier AI labs evaluate and audit their systems for biosecurity risk,” with four products: “Benchmarks,” “Audits” (“Independent, pre-deployment assessments of the biological risk a frontier model poses, run to feed directly into your responsible scaling policy”), “Red-teaming” and “Agentic KYC” (“Autonomous identity and legitimacy checks on applicants, so gene synthesis companies and frontier AI labs know exactly who they’re granting access to”) (LatchBio, Biosecurity). Unlike SecureBio, a nonprofit whose virology test sits in many labs’ cards, LatchBio is a venture-backed company with a data platform for biotech beside its evaluation line. Unlike Gray Swan and Trajectory Labs, what the labs print is its benchmark and, for xAI, its analysis; neither lab’s document describes LatchBio red-teaming that lab’s safeguards by hand.

Mandate & Origin

Latch Bio was founded in 2021 by Alfredo Andere, Kyle Giffin, and Kenny Workman, who met as undergraduates at the University of California, Berkeley. That is the company’s own seed announcement of 1 October 2021: “Latch Bio was founded in 2021 by Alfredo Andere, Kyle Giffin, and Kenny Workman who met as undergraduates at the University of California, Berkeley,” with Workman as “co-founder and Chief Technology Officer,” Giffin as “co-founder and Chief Operating Officer” and Andere as “co-founder and Chief Executive Officer” (Business Wire release, 1 Oct 2021, via BioSpace). Its manifesto: “The goal of The Latch Project is to build and disseminate the data infrastructure of the biocomputing revolution” (LatchBio, About). Its present mission line in job postings: “Our mission is to advance humanity’s ability to engineer biology. We approach this by understanding and improving AI’s application to biology across basic research, drug discovery, and biosecurity” (LatchBio job posting, Product Manager).

Legal entity. The terms of service name “LatchBio, Inc., with its principal place of business at 1800 Owens St, San Francisco, CA 94158” (LatchBio, Terms of Service). The paper that introduced its refusal benchmark gives the affiliation “LatchBio, San Francisco, CA” (arXiv:2607.05462).

Who runs it. A current posting lists “Founders: Alfredo, Kyle, Kenny” and a “Chief Product Officer, Hannah Le” (LatchBio job posting, Product Manager). The refusal paper lists Kenny Workman as a corresponding author with 15 co-authors from LatchBio and from American Wetware of Boston, among them Edwin H. Wintermute and Patrick M. Boyle (arXiv:2607.05462).

Scale, in its own words. “Latch works with 5,000 scientists, 150+ R&D labs, and many frontier labs to help develop and deploy the next generation of biological AI Agents” (LatchBio job posting, Research Partnerships Lead, 29 Aug 2026); “We’re roughly 95 people in San Francisco” (July) and “a team of roughly 100 people” (September) (LatchBio job posting, Proteomics; LatchBio job posting, Cell Therapy).

The acquisition. On 29 June 2026 its blog announced: “LatchBio has acquired TwentyTwo, a YC S25 team building AI infrastructure for biosecurity,” whose founders “will join Latch as Members of Technical Staff and lead Latch Biosecurity” (LatchBio blog).

Funding & Backers

The rounds, in the company’s own documents. The seed: “a $5 million seed funding round led by Lux Capital with participation from General Catalyst, Haystack, Fifty Years, and Asimov co-founder and CEO, Alec Nielsen, Ph.D.” (Business Wire release, 1 Oct 2021, via BioSpace). The Series A, in its postings: “$28M Series A from Lux Capital, Coatue, and General Catalyst” (LatchBio job posting, Manager, Special Projects, 3 Jul 2026). The current investor list: “backed by Lux Capital, Coatue, Fifty Years, Haystack, Jeff Dean, Alec Nielsen, and George Church” (LatchBio job posting, Proteomics). No round after the Series A, and no valuation, is stated on the company’s own pages or postings.

Customers. Its postings: “Frontier AI labs use our benchmarks to assess biological capabilities and safeguards. Our work appears in research announcements and model cards, including:” and then list Anthropic’s Fable 5.1 and Mythos 5.1 card, Opus 5.5 card, Anthropic’s “Advancing Claude in healthcare and life sciences,” xAI’s Grok 4.7 card and xAI’s “Biosecurity at the frontier” (LatchBio job posting, Product Manager). Its platform: “empowering over 5,000 scientists across 150+ R&D labs” (LatchBio job posting, AI Biologist, Metagenomics).

What the documents say about payment. No Anthropic or xAI document states a fee, a contract or a commission for LatchBio’s work. Anthropic’s Fable 5.1 card implies it runs LatchBio’s evaluations on its own models (“We did not run private benchmarks (such as LatchBio’s) on GPT or Gemini”); xAI’s post calls the analysis “independent” and says “The work of third-party evaluators such as LatchBio complements the evaluations we perform internally” (Anthropic, Fable 5.1; xAI, Biosecurity at the frontier).

Institutional Voice & Intent

The company writes in the register of a lab that measures and publishes. On 1 September 2026, the day Anthropic’s Fable 5.1 card and xAI’s biosecurity post appeared, its blog led with: “The currently-served version of Grok 4.6 is the most performant model we have tested on our biosecurity refusal benchmark, as measured by a weighted sum of disguised red-team refusals and routine, dual-use-adjacent research completions. This is a significant improvement from the earlier-tested versions of Grok 4.6,” and “Grok 4.6 was the only model that we tested that scored above 50% in both red-team refusal and routine answer rates, regardless of which harness we tested it with” (LatchBio blog, Testing Grok 4.6’s Enhanced Biology Safeguards, 1 Sep 2026). The same day, on Anthropic’s model: “the dominant trend is large number of refusals across a wide range of biological queries,” including “100% of all long-horizon tasks, and nearly all of tasks on our biosecurity capabilities and therapeutics benchmarks” (LatchBio blog, When It Answers, Fable 5.1 is Strong at Biology Reasoning, 1 Sep 2026).

• On the two labs’ safeguards, its paper. “Across 16 model-harness configurations, refusal rates ranged from 7% to 74% on Routine tasks and 1% to 62% on Red-Team tasks, with many configurations refusing legitimate Routine work at comparable or higher rates than concealed hazards. Refusals were most often triggered by provider API filters applied prior to agentic reasoning” (arXiv:2607.05462, 21 Jul 2026).
• On classifiers. “existing open-source guards do not flag red-team tasks effectively, block legitimate research, and often use incorrect rationales” (LatchBio blog, Open Source Classifiers Do Not Stop AI Bioweapon Generation, 7 Aug 2026).
• On its own tests. “Evaluations for agents in biology should not be stagnant, but rather continuously update as agentic capability and behavior changes over time”; after the update “The average model lost about 3.6 percentage points per benchmark” and “Anthropic models lost 3.1 percentage points, followed by OpenAI models losing 4.0, and SpaceXAI models losing 4.1” (LatchBio blog, Continuous Benchmarks for Biology to Defend Against Misalignment, 23 Sep 2026).
• On the stakes. “Within our lifetimes, AI will put the ability to engineer a pandemic within reach of a single bad actor”; “We cannot wall them off from the tools that augment their work” (LatchBio blog, 29 Jun 2026).

Stated intent: “Scaling AI in Biology Responsibly”; measure “whether your models meaningfully raise biological risk, and by how much,” without refusing the legitimate science.

Observed intent: Make its benchmarks the ones labs cite, sell the audits that sit next to them, and keep building the biotech data platform its postings describe. Its benchmark site publishes its own results for models from Anthropic, OpenAI, Google, Meta and Mistral.

Gap: The record shows two labs printing LatchBio’s numbers and a company that publishes unflattering ones about one of them. The two labs’ documents do not agree with each other on the same company’s output: the 1 September xAI post gives Grok 4.6 a 59.2% red-team refusal rate and a 62.1% average across harnesses, while the 21 September Grok 4.7 card’s table gives “Grok 4.6 (high)” 45.6% on Refusal with the Grok Build harness. The record does not show what either lab paid, whether LatchBio’s work for xAI was commissioned, or what either engagement lets LatchBio publish. The documents settle what was tested and printed. They do not settle the commercial terms.

Interest: LatchBio argues three positions on its own pages; it lodges none with a legislature, so what adoption “would pay” is read off its product page and postings.

• Refusal should be judged by whether a model understands intent, not by keywords or provider filters. Position, in xAI’s description of the benchmark’s design: “an agent that only reacts to words like “pathogen” or “toxin” will block the routine set and miss the red-team set, while one that inspects the data can tell the two apart”; and its blog’s “the Anthropic, OpenAI, Gemini, and xAI’s APIs all tended to refuse routine scientific queries more than red-team ones with latent biosecurity risk” (xAI; LatchBio blog, 7 Aug 2026). What it pays: BioSecBench-Refusal is the instrument that measures it, and the company sells the calibration work. Best case, LatchBio’s: the paper’s 7% to 74% spread on routine tasks, with refusals “most often triggered by provider API filters applied prior to agentic reasoning.” Convergence: xAI, whose post says “We gauge this risk as equally serious as the risk of aiding malicious use” and whose model the benchmark ranked first.
• Frontier labs should buy independent pre-deployment biology audits that feed their scaling policies. Position: “Independent, pre-deployment assessments of the biological risk a frontier model poses, run to feed directly into your responsible scaling policy” (LatchBio, Biosecurity). What it pays: “Request an audit” is the page’s button; a policy that makes outside audits a condition of release is demand for the product. Best case, LatchBio’s: “Latch builds the measurement layer for biosecurity: the benchmarks, audits, and red-teaming that tell you whether your models meaningfully raise biological risk, and by how much.” Convergence: Anthropic’s responsible scaling policy, whose Haiku 5.5 card now cross-references LatchBio’s suite “to assess relevant capabilities in bioinformatics,” and SecureBio and Gryphon Scientific, which sell biology evaluation to the same labs.
• Verify who gets access. Position: “Agentic KYC” so “gene synthesis companies and frontier AI labs know exactly who they’re granting access to” (LatchBio, Biosecurity). What it pays: a screening or access-control rule for synthesis providers and labs would create the market for the product. Best case, LatchBio’s: “Studying SARS-CoV-2 or Ebola to prepare for the next pandemic is very similar to what a bad actor would do to cause one,” so the line has to be drawn by who is asking (LatchBio blog, 29 Jun 2026). Convergence: the gene synthesis companies and frontier labs the product page names as its customers for the check.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

• Social accounts and feeds. X @latchbio, linked from latch.bio; feed: LatchBio blog and LatchBio on Substack.

Position in the Apparatus

LatchBio sits in the layer of the evaluator map that supplies the tests the labs run, as SecureBio and Irregular do, rather than the layer that attacks a model by hand. It differs from SecureBio, which is a nonprofit, in being a venture-backed company with an unrelated platform business, and it differs from Epoch AI and Artificial Analysis in that its best-known benchmark is of risk, not of capability alone.

• xAI: Grok 4.6 and Grok 4.7. Grok 4.6’s model card is dated 12 August 2026 and names no outside evaluator for biology; it says “We additionally provided an unrestricted configuration of Grok 4.6 to third-party evaluators” for cyber (xAI, Grok 4.6 model card). LatchBio’s blog on 13 August: “Following yesterday’s Grok 4.6 release, we ran it against our short-horizon biology tasks on benchmarks.bio” (LatchBio blog). On 1 September xAI published “Biosecurity at the frontier”: “Today, LatchBio published an independent analysis of Grok 4.6 on their biological capability and biological red-teaming benchmark suites,” and “On LatchBio’s BioSecBench-Refusal suite, Grok 4.6 was the strongest model tested at refusing disguised and hazardous tasks while still completing routine biological work. It was the only system to score above 50% on both measures” (xAI). On 21 September the Grok 4.7 post said the model topped “LatchBio’s biosafety benchmark at 62.4%,” and the card carries a section “LatchBio Capabilities v1.0” and a table where Refusal reads “Grok 4.6 (high)” 45.6%, “Grok 4.7 (xhigh)” 62.4%, “GPT-6 Astra (max)” 25.5% and “Opus 5 (max)” 31.8%, with “GPT-6 Astra and Opus 5 are LatchBio runs at max effort” (xAI, Grok 4.7; xAI, Grok 4.7 model card). The card’s own caveat: “Provider-side filters can dominate refusal behavior, and the restricted task sets characterize the tested checkpoint and harness rather than establishing general biological safety or reliability.”
• Anthropic: six cards and a post. Anthropic’s January 2026 life-sciences post charts “SpatialBench: Spatial biology analysis by LatchBio” for Opus 4.5 (Anthropic, Advancing Claude in healthcare and life sciences, 11 Jan 2026). The cards carry a section “LatchBio Bioinformatics,” “Developed by LatchBio”: Fable 5 and Mythos 5 (9 June 2026), Opus 5 (24 July), Fable 5.1 and Mythos 5.1 (1 September), Opus 5.5 (22 September), Sonnet 5.5 (28 September) and Haiku 5.5 (7 October), each reporting results on SpatialBench Verified and SingleCellBench (Anthropic, Fable 5 and Mythos 5; Anthropic, Opus 5; Anthropic, Fable 5.1; Anthropic, Opus 5.5; Anthropic, Sonnet 5.5; Anthropic, Haiku 5.5). The Haiku 5.5 card moves it into the threat-model evidence: “As of this model release, BioMysteryBench has been deprecated and removed from our panel of Life Sciences capabilities evaluations. In its place, we now cross-reference the LatchBio Bioinformatics suite to assess relevant capabilities in bioinformatics.” On Fable 5.1 the card says “Although many of these evaluations are not publicly released, we briefly describe each one below” and “We did not run private benchmarks (such as LatchBio’s) on GPT or Gemini.”
• Its site on Anthropic’s models. benchmarks.bio’s provenance note says of four Opus 5.5 and Fable 5.1 configurations: “the Claude API’s own classifier refused 1281 of their 1290 runs before the model produced a token, which scores as a refusal on both harnesses” (benchmarks.bio).
• Not named elsewhere. Meta’s Muse Spark report, Google DeepMind’s Gemini 4 Argon report and OpenAI’s GPT-5.5, GPT-5.6, GPT-6 Astra and GPT-6.1 Sol pages do not name LatchBio (Meta; Google DeepMind; OpenAI). benchmarks.bio nonetheless lists results for models from those labs, which LatchBio ran itself.

Actions & Leadership Choices

CONDUCT: NO CONDUCT ISSUE ON THE RECORD — A VENDOR OF BIOSECURITY AUDITS WHOSE BENCHMARK TWO LABS PUBLISH; ONE OF THEM IS THE LAB THE BENCHMARK RANKED FIRST. No wrongdoing is alleged. The documented facts: the company serves biotech customers by its own account and offers audits to labs; it was cited by xAI for a ranking that put xAI’s model first and by Anthropic as capability evidence; it publishes critical findings about Anthropic’s refusals and the open-source classifiers; and no document states what any lab paid.

What it did in 2026. It released fourteen benchmarks between February and October 2026 on its own site (OmicsBench and OncologyBench in October; MetagenomicsBench and the TxBench pair in September; BioSecBench-Function in August; VariantBench, BioSecBench-Surveillance and BioSecBench-Refusal in July) (LatchBio); acquired TwentyTwo to form Latch Biosecurity on 29 June; and published the BioSecBench-Refusal paper, whose public release is limited: “A public subset of BioSecBench-Refusal is available” and “107-evaluation set is held under restricted access” (arXiv:2607.05462). On 1 July 2026 DARPA’s Bio-Attribution Challenge results listed “TwentyTwo Bio (from Round 1)” with “Fastest Analysis” and “Fastest methods for executing data analysis” (DARPA, 1 Jul 2026); LatchBio’s security page claims the result: “Our algorithms took the fastest-analysis prize in DARPA’s Bio-Attribution Challenge.”

Alliance Map

These run at the same time. One venture-backed company is cited as evidence in a lab’s biological-risk panel, is the source of another lab’s headline safety number, publishes criticism of the first lab’s refusals, sells the audits that sit beside the benchmark, and holds a DARPA-recognized biosecurity team it bought two days before the award. Each tie is stated by the company or by a lab. The overlap is the finding.

Reach Assessment

Institutional: Its tests sit in two labs’ release documents, and in Anthropic’s case inside the capability evidence for a threshold under its scaling policy. Memetic: The public sees it through xAI’s launch posts, which lead with its number, and through its own blog, which publishes the same benchmark’s less welcome findings. Civilizational: A benchmark a lab uses to say a model is calibrated is a benchmark that sets what calibrated means; if outside biology audits become a required part of a release, a company that already holds one lab’s evidence panel and another’s launch claim holds a seat in that decision; today it holds it by citation.

Score basis. Atk 6 is reach: six Anthropic cards and Anthropic’s Haiku 5.5 threat-model language, and xAI’s two posts and card (Anthropic, Haiku 5.5; xAI, Grok 4.7 model card). It sits level with Vals AI (6/4/5, hazard 50) and Andon Labs (6/5/6, hazard 58), whose benchmarks and tests are also printed in lab cards, and one under Trajectory Labs (7/4/5, hazard 53), named by three labs. The points under 7 are that two labs name it, that Meta, Google DeepMind and OpenAI do not, and that no statute obliges a lab to run it. Def 5 is funding insulation: its backers are venture funds and individuals, not labs, but it sells audits to labs and two labs cite it (LatchBio, Biosecurity); that is level with Faculty (8/5/6), 10a Labs (5/5/4) and Andon Labs (5), one over Trajectory Labs (4). Hp 6 is entrenchment: five years of operation, about a hundred staff and a platform for “5,000 scientists,” and a place in Anthropic’s life-sciences panel in place of BioMysteryBench; level with Andon Labs and Faculty (6), one over Trajectory Labs (5). Hazard 58 sits level with Andon Labs and one under Artificial Analysis (8/4/6, hazard 60), which has more reach and less insulation.

Sources: LatchBio; Biosecurity — LatchBio; About Us — LatchBio; benchmarks.bio; Terms of Service — LatchBio; Testing Grok 4.6’s Enhanced Biology Safeguards — LatchBio; Grok 4.6 is a Frontier Biology Model — LatchBio; When It Answers, Fable 5.1 is Strong at Biology Reasoning — LatchBio; Open Source Classifiers Do Not Stop AI Bioweapon Generation — LatchBio; Continuous Benchmarks for Biology to Defend Against Misalignment — LatchBio; LatchBio acquires TwentyTwo to form Latch Biosecurity — LatchBio; BioSecBench-Refusal — arXiv:2607.05462; Biosecurity at the frontier — xAI; Introducing Grok 4.7 — xAI; Grok 4.7 model card — xAI; Grok 4.6 model card — xAI; Claude Haiku 5.5 System Card — Anthropic; Claude Sonnet 5.5 System Card — Anthropic; Claude Opus 5.5 System Card — Anthropic; Claude Fable 5.1 and Mythos 5.1 System Card — Anthropic; Claude Opus 5 System Card — Anthropic; Claude Fable 5 and Mythos 5 System Card — Anthropic; Advancing Claude in healthcare and life sciences — Anthropic; Bio-Attribution Challenge yields tools to define biothreat origins at speed, scale — DARPA; Latch Bio Emerges from Stealth with Seed Funding Round Led by Lux Capital — BioSpace; LatchBio jobs — Ashby; Muse Spark Safety and Preparedness Report — Meta; Gemini 4 Argon safety report — Google DeepMind; GPT-6 Astra — OpenAI.